A genetic algorithm with memory for optimal design of laminated sandwich composite panels

A genetic algorithm with memory for optimal design of laminated sandwich composite panels

Composite Structures 58 (2002) 513–520 www.elsevier.com/locate/compstruct A genetic algorithm with memory for optimal design of laminated sandwich co...

175KB Sizes 0 Downloads 101 Views

Composite Structures 58 (2002) 513–520 www.elsevier.com/locate/compstruct

A genetic algorithm with memory for optimal design of laminated sandwich composite panels Vladimir B. Gantovnik a, Zafer G€ urdal a, Layne T. Watson a b

b,*

Department of Engineering Science and Mechanics, Virginia Polytechnic Institute and State University, Blacksburg, VA 24061-0219, USA Departments of Computer Science and Mathematics, Virginia Polytechnic Institute and State University, 630 McBryde Hall, Blacksburg, VA 24061-0106, USA

Abstract This paper is concerned with augmenting genetic algorithms (GAs) to include memory for continuous variables, and applying this to stacking sequence design of laminated sandwich composite panels that involves both discrete variables and a continuous design variable. The term ‘‘memory’’ implies preserving data from previously analyzed designs. A balanced binary tree with nodes corresponding to discrete designs renders efficient access to the memory. For those discrete designs that occur frequently, an evolving database of continuous variable values is used to construct a spline approximation to the fitness as a function of the single continuous variable. The approximation is then used to decide when to retrieve the fitness function value from the spline and when to do an exact analysis to add a new data point for the spline. With the spline approximation in place, it is also possible to use the best solution of the approximation as a local improvement during the optimization process. The demonstration problem chosen is the stacking sequence optimization of a sandwich plate with composite face sheets for weight minimization subject to strength and buckling constraints. Comparisons are made between the cases with and without the binary tree and spline interpolation added to a standard GA. Reduced computational cost and increased performance index of a GA with these changes are demonstrated. Ó 2002 Published by Elsevier Science Ltd. Keywords: Genetic algorithm; Composite panel structure; Spline approximation

1. Introduction Traditionally, the problem of composite laminate stacking sequence optimization has been defined as a continuous design problem and solved using gradientbased techniques [1]. However, because of manufacturing considerations, the orientations of the fibers in plies are typically confined within a discrete set, for example, (0°, þ45°, 45°, 90°). The stacking sequence optimization in this case could be formulated as an integer-programming problem [2]. Recent studies have shown that genetic algorithms (GAs) are highly suitable for the solution of the composite laminate design problems with discrete design variables [3,4]. The solution of such problems by GA is possible, because the method does not require gradient or Hessian information. A GA is a powerful technique for search and optimization problems with discrete variables, and is therefore particularly *

Corresponding author. Tel.: +1-540-231-7540; fax: +1-540-2316075. E-mail address: [email protected] (L.T. Watson).

useful for optimization of composite laminates. However, to reach an optimal solution with a high degree of confidence, it typically requires a large number of analyses during the optimization search. Performance of GAs is even more of an issue for problems that include continuous variables. Several studies have concentrated on improving the reliability and efficiency of GAs. Hybrid algorithms formed by the combination of a GA with local search methods provide increased performance when compared to a GA with a discrete encoding of real numbers or local search alone [5]. In order to reduce the computational cost, two of the authors earlier used local improvements and memory so that information from previously analyzed design points is utilized during a search [6,7]. In the first method a memory binary tree was employed to store pertinent information about laminate designs that have already been analyzed [6]. After the creation of a new population of designs, the tree structure is searched for either a design with identical stacking sequence or similar performance, such as a laminate with identical in-plane strains. Depending on

0263-8223/02/$ - see front matter Ó 2002 Published by Elsevier Science Ltd. PII: S 0 2 6 3 - 8 2 2 3 ( 0 2 ) 0 0 1 2 8 - 9

514

V.B. Gantovnik et al. / Composite Structures 58 (2002) 513–520

the kind of information that can be retrieved from the tree, the analysis for a given laminate may be significantly reduced or may not be required at all. The second method is called local improvement [7]. This technique was applied to the problem of maximizing the buckling load of a rectangular laminated composite plate. The information about previously analyzed designs is used to construct an approximation to buckling load in the neighborhood of each member of the population of designs. After that the approximations are used to search for improved designs in small discrete spaces around nominal designs. These two methods demonstrated substantial improvements in computational efficiency for purely discrete optimization problems. The implementation, however, was not suitable for handling continuous design variables. The objective of the present work is to find a suitable algorithm for a GA with memory that can work with discrete and continuous variables simultaneously. A local memory for the continuous part of the design space based on spline approximation is proposed for problems with a single continuous variable. The efficiency of the proposed spline based procedure, as well as the use of memory for a GA that can handle continuous variables, are investigated for the weight optimization of a sandwich plate with composite face sheets subjected to strength and buckling constraints.

Fig. 1. The structure of a population.

the population, and there is just one laminate chromosome with one material type. Here only the ply orientation genes are used for the stacking sequence definition and a single geometry gene is used for the single continuous variable.

3. Binary tree memory 2. Genetic algorithm package A state-of-the-art Fortran 90 GA framework that was designed in an earlier research effort was used for the composite laminate structure design [8]. This framework includes a module, encapsulating GA data structures, and a package of GA operators. The module and the package of operators result in what we call a standard GA. The proposed algorithm is incorporated within the GA framework as a sample test program that illustrates performance of the binary tree memory and spline interpolation. An integer alphabet is used to code ply genes. It should be pointed out that the continuous variables had already been implemented in the GA data structure as geometry chromosomes, represented directly as real numbers and not discrete binary approximations. Fig. 1 from [8] shows the data structure of a population for the most general case of composite structure design. This data structure supports advanced features such as migration between independently evolving subpopulations, use of multiple materials for designing hybrid laminates, and multiple laminated substructures represented by multiple laminate chromosomes. In general, the laminate chromosomes are for discrete variables, and the geometry chromosomes are for continuous variables. Migration is not used here, however, so there is just one subpopulation identical to

A binary tree is a linked list structure in which each node may point to up to two other nodes. In a binary search tree, each left pointer points to nodes containing elements that are smaller than the element in the current node; each right pointer points to nodes containing elements that are greater than the element in the current node, as shown in Fig. 2. A binary tree has several properties of great practical value, one of which is that the data can be retrieved, modified, and inserted relatively quickly. If the tree is perfectly balanced, the cost of inserting of an element in a tree with n nodes is proportional to log2 n steps, and rebalancing the tree after an insertion may take as little as several steps, but at most takes log2 n steps. Thus, the total time is of the order of log2 n [9]. In the standard GA, a new population may contain designs that have already been encountered in the

Fig. 2. An example of a binary tree.

V.B. Gantovnik et al. / Composite Structures 58 (2002) 513–520

Fig. 3. Evaluation of fitness function using binary tree.

previous generations, especially towards the end of the optimization process. The memory procedure eliminates the possibility of repeating an analysis that could be expensive. The binary tree, keyed on the discrete design string encodings, is used to store data pertinent to the design such as the design string and the fitness and constraint functions. In Fig. 2, for example, the integers in the boxes represent design string encodings, used as keys for comparison to traverse the tree. The boxes (binary tree nodes) would also contain fitness and constraint function values (not shown) for that discrete design. Fig. 3 shows the pseudocode for the fitness function evaluation with the aid of the binary tree. After a new generation of designs is created by the genetic operations, the binary tree is searched for each new design. If the design is found, the fitness value is retrieved from the binary tree without conducting an analysis. Otherwise, the fitness is obtained based on an exact analysis. This new design and its data are then inserted in the tree as a new node.

515

the continuous values and the associated fitness function. That is, each node contains a real array that stores the continuous variable and its corresponding value of the fitness function. In order for the memory to be functional, it is necessary to have accumulated enough designs with different continuous values for a particular discrete design part. Naturally, not all the nodes will have more than a few designs with different continuous values. However, it is possible that as the evolution progresses good discrete parts will start appearing again and again with different continuous values. In this case, one will be able to construct a good quality spline interpolant to the data. In addition to building a spline approximation it is important to develop a scheme in which accuracy of the spline approximation at a new continuous point may be assessed, so that a decision may be made either to accept the approximation or perform exact function evaluation. In the proposed approach for all the design points with identical discrete variables and different continuous variable values, a measure of distance within which a good quality approximation may be obtained is used to make the decision. Based on the stored information, the procedure described by pseudocode in Fig. 4 decides whether an exact analysis should be performed and stored in the tree, or to extract an approximation of the fitness value without an exact analysis. For the description of the pseudocode, let v be a discrete vector variable, x a scalar real variable, Uðv; xÞ the fitness of the individual defined by ðv; xÞ, and d a real number representing the radius of an interval within which a good quality approximation is possible. Each

4. Spline interpolation The procedure described above works well for designs that are completely described by discrete strings. In case of designs that include a continuous variable, the solution is more complicated. If the continuous variable is also discretized into a fine discrete set, the possibility of creating a child design that has the same discrete and continuous parts as one of the earlier designs diminishes substantially. In the worst case, if the continuous design variable is represented as a real number, which is the approach used by most recent research work, it may be impossible to create a child design that has the exact same real part as one of the parents, rendering the binary tree memory useless. The main idea of the approach proposed in this paper is to construct approximations for the fitness function as a function of the continuous variable using a spline function fitted to the historical data, and interpolate from the stored data whenever possible. The memory in this case consists of two parts. A binary tree, which consists of the nodes that have different discrete parts of the design, and a storage part at each node that keeps

Fig. 4. Evaluation of fitness function using binary tree and spline.

516

V.B. Gantovnik et al. / Composite Structures 58 (2002) 513–520

node in the binary tree memory structure records the n data ðv; fxi ; Uðv; xi Þ; di gi¼1 Þ, where n is the number of times the design is evaluated with different values of the real design variable corresponding to the same discrete part. For the first three values of the continuous variable at a given discrete branch, the points are automatically evaluated and entered into the spline database with radius values d of zero. Since the initial spline points will have zero radius, many of the additional points x will not satisfy the proximity condition max1 6 i 6 n fdi  jx xi jg P 0 initially, and therefore their fitness Uðv; xÞ will be computed exactly. It is possible to add the new points either with a zero radius or a positive computed value for the radius. If the actual value of the fitness Uðv; xÞ and the spline approximation SðxÞ are within acceptable error limit , then the radius value d for the new point and the closest point xk to the new point are both changed to the distance between the two points, d ¼ minfd0 ; jx  xk jg, and the new point is added to the spline database with this computed value. Otherwise, the new point is added to the database with a zero radius value (see Fig. 4, for the pseudocode). The parameter d0 is an upper bound for the distance d that protects the algorithm against large errors between exact and spline function values. The approximation of the fitness function is based on spline interpolation of order less than or equal to four with the ‘‘not-a-knot’’ boundary condition. The ‘‘nota-knot’’ condition is used when no information about the end point derivatives of an approximated function is known. The subroutine CUBSPL from the Fortran 90 version of Carl de Boor’s cubic spline package was used to construct the spline approximations [10].

5. Local improvement Local improvement is essentially an addition to improve the performance of the GA with spline interpolation. The value of the continuous variable at a given discrete node is either randomly assigned or obtained through a crossover operation. If an explicit spline approximation is available at a given node, it is possible to use a spline interpolant to generate a good candidate for the continuous variable for the next point at that node rather than depend on random action from the crossover operator. That is, after construction of an initial approximation SðxÞ of Uðv; xÞ based on the objective function values obtained at the design sites, one can easily find the point x that minimizes the spline function SðxÞ in the range ½x1 ; xn . This optimal x value is stored at the discrete node in addition to the rest of the spline database. If, in future generations, a discrete node that has a stored x value is reached through the crossover operation on the discrete part v of the design, then, rather than performing crossover on the real part ðv; x Þ is

used as the child design for the next generation. This child design will then be treated like the other new designs in the child population and will be checked if an approximation to it can be used without exactly evaluating it.

6. Analysis The optimum design of a honeycomb sandwich construction for strength and buckling constraints is treated in this paper. The core depth, which is a real variable, the ply orientation angles and the number of layers, which are discrete variables, are taken as the design variables. Consider a honeycomb sandwich plate with length a, width b, face ply thickness tp , core depth th , core cell wall thickness tc , and cell size diameter td . When such a structure is subjected to biaxial compression, there are several modes of sandwich plate failure. One obvious way is overstressing. The strength of faces is determined by the maximum strain criterion. The plate is assumed to fail if any of the ply strains, fi1 ; i2 ; ci12 gni¼1 , exceeds its allowable value. The critical strength failure load factor is defined as      c ks ¼ min min 1i ; 2i ; 12 ; ð1Þ 16i6n 1 1 ci12 where f 1 ;  2 ; c 12 g are the ultimate allowable strains. The strength constraint is represented by Gs ¼ ks  1:0 P 0:

ð2Þ

Before overstressing, it is also possible that the structure can buckle. The reliability for the optimum design of the honeycomb sandwich construction is enhanced by adding several buckling constraints. Three modes of instability are considered as buckling constraints: overall buckling, face wrinkling, and face dimpling. The overall buckling constraint for a honeycomb sandwich panel is represented by   P P Gb ¼ min cr ; cr  1:0 P 0; ð3Þ Nx Ny where Pcr is the lowest value of the critical buckling load. The computation of the buckling load is based on the mathematical model proposed by Hwu and Hu [11]. A closed-form solution of the buckling load was used for the composite sandwich plate consisting of crossply symmetric laminate faces with all edges simply supported: Pcr ¼

ð1 þ gÞP0 2

2

½1 þ kða=bÞ ðn=mÞ ð1 þ gx þ gy þ g0 Þ

;

ð4Þ

where g¼

  p2 D 1  m 2 1  n 2 þ ; Gyz a Gxz b th

g0 ¼

D P0 m2 n2 ; a2 th2 Gxz Gyz

V.B. Gantovnik et al. / Composite Structures 58 (2002) 513–520

    n 2  p2 m 2 gx ¼ D11 þ D66 ; a b th Gxz     n 2  p2 m 2 gy ¼ D66 þ D22 ; a b th Gyz     mn 2 p2 a2 m 4 P0 ¼ 2 D11 þ 2ðD12 þ 2D66 Þ a ab m  n 4  þ D22 ; b   m 4 p2 a2 D P0 ¼ 2 D11 D66 þ ðD11 D22  D12 ðD12 a m  mn 2  n 4  þ 2D66 ÞÞ þ D22 D66 ; ab b k ¼ Ny =Nx is the loading ratio; m and n are the number of buckling half-waves within the deformed surface; Dij are the bending stiffnesses of the laminate with respect to the midsurface of the core; Gxz and Gyz are the transverse shear moduli of the core material in the ðx  zÞ and ðy  zÞ planes. Face wrinkling may occur across many cells of the honeycomb core. The face wrinkling constraint can be expressed as 2 3  fw  fw N N y 5 Gfw ¼ min 4 x ;  1:0 P 0; ð5Þ Nx Ny where  fw ¼ N x

 fw N y

Nxfw Nyfw

"

1 Nx ð1 þ Gs Þ

2  4 ¼



3

1 Ny ð1 þ Gs Þ

þ

1 Nxfw

þ

1 Nyfw

3

3 #1=3 ; !3 31=3 5 ;

  1=3 D212 ¼ D11  ; Ec Gxz D22   1=3 D212 ¼ D22  : Ec Gyz D11

In honeycomb core sandwiches, a third type of instability, which is generally referred to as the face dimpling, may occur because the face sheet over one cell of the honeycomb can buckle as a small plate supported by the cell walls. The face dimpling constraint can be written as 2 3 fd fd N   N y 5 Gfd ¼ min 4 x ;  1:0 P 0; ð6Þ Nx Ny

 fd N y

Nxfd Nyfd

"

1 Nx ð1 þ Gs Þ



3 þ

1 Nxfd

3 #1=3 ;



1 ¼4 Ny ð1 þ Gs Þ

3 þ

1 Nyfd

!3 31=3 5 ;

  3=2 9 D212 tf ¼ 2 D11  ; tf D22 td   3=2 9 D2 tf ¼ 2 D22  12 : tf D11 td

7. Optimization problem The optimization problem can be formulated as finding the stacking sequences of the face sheet and the core thickness in order to minimize the weight W of the panel. The set of design variables is expressed as a vector ~t ¼ ðh1 ; . . . ; hn ; th Þ, where n is an implicit design variable dictated by the number of layers in the face sheet stacking sequence. The optimization problem with displacement and buckling constraints can be expressed as min W ð~tÞ;

ð7Þ

~t

such that Gs ð~tÞ P 0 (strength constraint), Gb ð~tÞ P 0 (overall buckling), Gfw ð~tÞ P 0 (face wrinkling), Gfd ð~tÞ P 0 (face dimpling), th 2 ½thl ; thu (core thickness), hi 2 f0°; 45°; 90°g; i ¼ 1; . . . ; n (ply angles), where W ð~tÞ ¼ ½ð2tf Þqf þ th ð8=3Þðtc =td Þqc ðabÞ is the structural weight of the plate; qc and qf are material densities tf ¼ Pn=2 i of the core and face ply, l respectively; u i¼1 tply is the total face thickness; th and th are lower and upper bounds of the core thickness, respectively; hi is the orientation of the ith ply; n is the total number of plies. The critical constraint is defined as Gcr ð~tÞ ¼ minfGs ; Gb ; Gfw ; Gfd g;

ð8Þ

and the constrained optimization problem is transformed into an unconstrained maximization problem for the GA. This is done by using penalty parameters. The fitness function U to be maximized is defined as Uð~tÞ ¼

where  fd ¼ N x

2

517



W ð~tÞ þ Gcr d; p W ð~tÞð1  Gcr Þ ;

Gcr P 0; Gcr < 0;

ð9Þ

where d and p are bonus and penalty parameters, respectively.

518

V.B. Gantovnik et al. / Composite Structures 58 (2002) 513–520

8. Results

8.1. Effect of binary tree

A T300/N5208 graphite-epoxy unidirectional ply is used for the laminate face construction. The core material is HFT fiberglass reinforced phenolic honeycomb. The material properties of the face and the core are listed in Table 1. The panel has the longitudinal and lateral dimensions of a ¼ 20 in. and b ¼ 10 in., respectively. The allowed range of the core thickness is from 0.1 to 1.5 in. The GA was applied for one load case with Nx ¼ 8000:0 lb/in. and Ny ¼ 0:0 lb/in. The GA stopping condition is either a limit on the total number of function evaluations conducted by the standard GA, or when there is no change in the fitness function between generations, whichever occurs first. For all the results here (other than Table 2, for which a limit of 125,000 was used as the stopping condition), the limit was set large enough that it was never reached.

In order to demonstrate the efficiency that can be achieved through the use of a discrete memory, first consider the problem with the core thickness variable fixed to a constant value. The best known global optimal design for the specified load case obtained by the standard GA is presented in Table 2. The table gives the average number from ten runs of individuals ð ni Þ, the thickness ðtÞ, ply orientations ðvÞ, and fitness function value ðUÞ. This design was obtained with the constant core thickness, th ¼ 1:0 in, in about 125,000 function evaluations. The efficiency of the binary tree for the fitness function evaluation is shown in Table 3. This table shows the averages from ten runs for the number of individuals (total number of attempts to evaluate fitness function) ðni Þ, the number of nodes in the binary tree ðne Þ, the number of repeated designs ðnr Þ, savings in terms of amount of fitness evaluations ðnÞ, ply orientations ðvÞ, and fitness function value ðUÞ. The chromosome length is 20. It is evident from this table that 40–50% of the analyses can be avoided by using the GA with memory. This reduction of the total number of analyses agrees well with the results from the previous studies of the GA with memory [6,7].

Table 1 Material properties Face place

Core

E1 ¼ 20:5  106 psi E2 ¼ 1:65  106 psi G12 ¼ 0:75  106 psi m12 ¼ 0:32 tply ¼ 0:006 in. qf ¼ 0:0055 lbs/in.3

Ec ¼ 23; 000 psi Gxz ¼ 19; 000 psi Gyz ¼ 7500 psi td ¼ 0:125 in. tc ¼ 1:0 in. qc ¼ 0:00174 lbs/in.3

8.2. Effect of spline interpolation

Table 2 Optimal design ni 

t

v

U

125,000

1.0

1111111

1.3616

Table 3 The efficiency of the binary tree ni 

ne 

nr

n (%)

v

U

84,572

43,937

40,635

47.0

1111111

1.3616

The results of this experiment were obtained with the core thickness varying in the range [0.1, 1.5] in. The best known optimal design is presented in Table 4. The design was obtained using a standard GA with a large number of iterations requiring over a million function evaluations. The performance of the GA with the spline interpolation is presented in Table 5, which shows averages from ten runs. Table 5 shows the best design after ni attempts to evaluate the fitness function, ns accepted spline values, and with  ¼ 0:005 and d0 ¼ 0:01. The chromosome length is 7. The best known discrete vector v is f1 1 1 1 1 1 1g. The maximum absolute error dmax and the mean absolute error dav are defined as dmax ¼ max jUðxi Þ  Sðxi Þj; 1 6 i 6 ns

Table 4 Optimal design

dav ¼

ni 

t

v

U

1,181,040

0.4220

1111111

1.1882

ns 1 X jUðxi Þ  Sðxi Þj; ns i¼1

ð10Þ

where ns is the total number of acceptable spline evaluations. These errors are computed every time that the

Table 5 The efficiency of the spline interpolation  ni

ns

n (%)

t

U

dmax

dav

969,011

896,487

71.3

0.4220

1.1882

0.11978

0.00006

V.B. Gantovnik et al. / Composite Structures 58 (2002) 513–520

519

algorithm decides to extract an approximation of the fitness value without an exact analysis. The data show the cost of the GA with a continuous variable could be reduced up to 70% relative to the standard GA by using the spline local approximation procedure. For the problem considered, the computation of the fitness function is not very expensive in terms of CPU time. However, this procedure has great potential in problems with expensive objective functions. Table 6 shows he maximum absolute error dmax and the mean absolute error dav for different values of the

parameters  and d0 . Table 7 shows the savings ðnÞ in terms of number of fitness evaluations for different values of the parameters  and d0 . Table 6 shows the expected trends; dav decreases as either  or d0 decreases, but dmax is essentially random, being determined early in the iterations. Note that the average approximation error dav is quite acceptable, several orders of magnitude less than  and d0 . Table 7 shows that the saving are significant, but of course these percentages will be smaller for larger chromosomes.

Table 6 The errors of the spline interpolation ðdmax ; dav Þ as a function of the parameters  and d0

8.3. Effect of local improvement

 0.005 0.01 0.02 0.04

d0 0.005

0.01

0.02

0.03434, 0.00003 0.06142, 0.00004 0.04859, 0.00005 0.05555, 0.00012

0.11978, 0.00006 0.09826, 0.00009 0.09676, 0.00012 0.10540, 0.00024

0.20111, 0.00022 0.23898, 0.00024 0.32709, 0.00030 0.34188, 0.00042

Table 7 The percent savings ðnÞ as a function of the parameters  and d0  0.005 0.01 0.02 0.04

d0 0.005

0.01

0.02

50.5 59.6 63.0 64.9

70.8 72.1 76.2 77.8

75.7 78.5 80.8 86.4

The fitness as a function of the natural logarithm of the number of exact fitness evaluations is plotted in Fig. 5. Since the analysis is a finite deterministic computation, involving no iterative algorithms, computation time corresponds exactly to number of function evaluations, which is shown in Fig. 5. If iterative processes were involved, then time would be more meaningful for the abscissa in the figure. It can be observed that the algorithm with local improvement converges faster in terms of number of fitness function evaluations than the one with spline approximation. The algorithm with only spline interpolation demonstrates good convergence in comparison with the standard GA, and noticeably decreases the number of exact analyses. After about 20,000 exact function evaluations the fitness values of the standard GA, the GA with spline interpolation, and the GA with local improvement are U ¼ 1:2095, U ¼ 1:2031, and U ¼ 1:1882, respectively.

Fig. 5. Comparison of the performance (fitness vs. number of exact analyses).

520

V.B. Gantovnik et al. / Composite Structures 58 (2002) 513–520

9. Conclusions A GA with memory along with spline interpolation was applied to the problem of weight minimization of a laminated sandwich composite panel with mixed discrete face sheet design variables and a continuous thickness design variable. The use of memory based on binary tree for discrete part of the design variables, and spline approximation for a continuous variable avoids repeating analyses of previously encountered designs. Moreover, it is also demonstrated that the spline approximation can be used to provide local improvement during the search and further reduce the number of exact function evaluations required to reach an improved solution. For FEM based analyses/simulations in which the computational cost is not trivial, the percent savings demonstrated in Tables 3, 5 and 7 will result in substantial computational savings. Even if less sophisticated analysis than FEM is used, the use of the binary tree will save time for all but extremely trivial function evaluations. Future work on this procedure might include development of a robust algorithm that is capable of handling several continuous variables using multivariate approximation techniques.

Acknowledgements The authors are indebted to Christine AndersonCook for suggesting improvements in the approximation algorithm. This research was supported in part by Air Force Office of Scientific Research grant F49620-99-

1-0128 and National Science Foundation grant DMS9625968. References [1] Schmit LA, Farshi B. Optimum design of laminated fiber composite plates. International Journal of Numerical Methods in Engineering 1977;11:623–40. [2] Haftka RT, Walsh JL. Stacking sequence optimization for buckling of laminated plates by integer programming. AIAA Journal 1992;3(3):814–9. [3] Nagendra S, Haftka RT, G€ urdal Z. Stacking sequence optimization of simply supported laminates with stability and strain constraints. AIAA Journal 1992;30(8):2132–7. [4] LeRiche R, Haftka RT. Optimization of laminate stacking sequence for buckling load maximization by genetic algorithm. AIAA Journal 1993;31(5):951–6. [5] Seront G, Bersini H. A new GA-local search hybrid for optimization based on multilevel single linkage clustering. In: Proceedings Genetic and Evolutionary Computation Conference (GECCO-2000) Las Vegas, Nevada, 10–12 July 2000. [6] Kogiso N, Watson LT, G€ urdal Z, Haftka RT. Genetic algorithms with local improvement for composite laminate design. Structural Optimization 1994;7(4):207–18. [7] Kogiso N, Watson LT, G€ urdal Z, Haftka RT, Nagendra S. Design of composite laminates by a genetic algorithm with memory. Mechanics of Composite Materials and Structures 1994;1(1):95–117. [8] McMahon MT, Watson LT, Soremekun GA, G€ urdal Z, Haftka RT. A Fortran 90 genetic algorithm module for composite laminate structure design. Engineering with Computers 1998;14: 260–73. [9] Vowels RA. Algorithms and data structures in F and Fortran. Tucson, Arizona: Unicomp, Inc.; 1998. [10] de Boor C. A practical guide to splines. Berlin: Springer-Verlag; 1978. [11] Moh J, Hwu C. Optimization for buckling of composite sandwich plates. AIAA Journal 1997;35(5):863–8.